Department of Computer, Control and Management Engineering

The epidemiology and clinical characteristics of COVID-19 evolved due to new SARS-CoV-2 variants of concern (VOCs). The Omicron VOC’s higher transmissibility increased paediatric COVID-19 cases and hospital admissions. Most research during the Omicron period has focused on hospitalised cases, leaving a gap in understanding the disease’s evolution in community settings. This study targets children with mild to moderate COVID-19 during pre-Omicron and Omicron periods. It aims to identify patterns in COVID-19 morbidity by clustering individuals based on symptom similarities and duration of symptoms and develop a machine-learning tool to classify new cases into risk groups. We propose a data-driven approach to explore changes in COVID-19 characteristics by analysing data from 581 children and adolescents collected within a paediatric cohort at the University Hospital of Padua. First, we apply an unsupervised machine-learning algorithm to cluster individuals into groups. Second, we classify new patient risk groups using a random forest classifier model based on sociodemographic information, pre existing medical conditions, vaccination status and the VOC as predictive variables. Third, we explore the key features influencing the classification through the SHapley Additive exPlanations. The unsupervised clustering identified three severity risk profile groups. Cluster 0 (mildest) had an average of 1.2 symptoms (95% CI 0.0 to 5.0) and mean symptom duration of 1.26 days (95%CI 0.0 to 9.0), cluster 1 had 2.27 symptoms (95% CI 1.0 to 6.0) lasting 3.47 days (95% CI 1.0 to 12.0), while cluster 2 (strongest symptom expression) exhibited 3.41 symptoms (95% CI 2.0 to 7.0) over 5.52 days (95% CI 0.0 to 16.0). Feature importance analysis showed that age was the most important predictor, followed by the variant of infection, influenza vaccination and the presence of comorbidities. The analysis revealed that younger children, unvaccinated individuals, those infected with Omicron and those with comorbidities were at higher risk of experiencing a greater number and longer duration of symptoms. Our classification model has the potential to provide clinicians with insights into the children’s risk profile of COVID-19 using readily available data. This approach can support public health by clarifying disease burden and improving patient care strategies. Furthermore, it underscores the importance of integrating risk classification models to monitor and manage infectious diseases.